AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

Latent Space Newsletter: Reflection Beam, OpenAI vs. Anthropic, and AI Research

· Latent Space

The Latent Space newsletter discusses the launch of Reflection's Beam, a 501B-total / 23B-active MoE model for coding, agentic, and scientific work, trained from scratch with full weights under Apache 2.0. It also features a detailed comparison of subscription value between Anthropic and OpenAI, with SemiAnalysis finding Claude subscriptions deliver significantly more API-equivalent value. The newsletter further delves into advancements in agent harnesses, RL environments, and developer tooling, including Hugging Face's multi-harness RL and Cognition's Devin "Dreaming" for agent memory. Finally, it covers various AI research topics such as verification over sampling, context management, self-improving harnesses, and optimization techniques like Dust and LOOM.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.